Speedrun Skill

Comprehensive optimization hunting that dispatches parallel agents to find performance bottlenecks across all layers, then applies fixes with mandatory verification.

The Iron Law

NO OPTIMIZATION CLAIM WITHOUT MEASUREMENT.

Every optimization must have before/after metrics. Unverified "improvements" are worthless.

ROI & Priority Rubric

All hunters use the same priority classification:

Priority	Impact	Fix Effort
P0	>50% improvement or >100KB savings	Trivial (1-2 lines)
P1	20-50% improvement or 50-100KB savings	Straightforward
P2	10-20% improvement or 10-50KB savings	Moderate
P3	<10% improvement or <10KB savings	Any

Confidence:

• HIGH: Direct evidence (measured file size, query count, profiler output)
• MEDIUM: Pattern match — savings estimated from code structure
• LOW: Theoretical — no direct measurement possible

Fix order: P0 first. Never spend time on P3 while P0 exists.

Quick Reference

Hunter	Profile	Focus	Color
bundle-hunter	OBSESSIVE	File sizes, tree-shaking, code splitting	yellow
complexity-hunter	UNCOMFORTABLE	Cyclomatic complexity, nesting, function length	purple
dead-code-hunter	RUTHLESS	Unused exports, orphan files, unreachable paths	gray
dependency-hunter	WARY	Bloat, duplicates, outdated, vulnerabilities	magenta
build-hunter	IMPATIENT	Compilation time, caching, parallelization	orange
algorithm-hunter	IMPATIENT	Big O, data structures, repeated computation	red
query-hunter	METICULOUS	N+1, missing indexes, slow queries	green
web-vitals-hunter	OBSESSIVE	LCP, CLS, INP, TTFB	blue
memory-hunter	PARANOID	Leaks, allocation patterns, GC pressure	red
image-hunter	METICULOUS	Formats, compression, lazy loading	cyan

Workflow

1. Measure baseline - Gather current metrics (bundle size, build time, complexity scores)
2. Detect stack - Identify framework, build tooling, database, image handling
3. Dispatch applicable hunters in PARALLEL (stack-gated — skip irrelevant hunters)
4. Collect findings - Aggregate with P0-P3 priority and confidence ratings
5. Write SPEEDRUN.md - Prioritized optimization report
6. Apply fixes - P0/P1 safe fixes auto-apply, complex fixes require approval
7. Verify improvements - Re-measure and report deltas

Stack-Gated Dispatch

Always dispatch: bundle, complexity, dead-code, dependency, algorithm

Dispatch if applicable:

• build-hunter — has build config (webpack/vite/rollup/turbo)
• query-hunter — has database (Supabase/Prisma/SQL/Mongoose)
• web-vitals-hunter — has frontend (React/Vue/HTML)
• memory-hunter — has server runtime (Node/Python/Go)
• image-hunter — has image assets (public/, static/, assets/)

Fix Tiers

Auto-Apply (Safe)

• Remove unused imports/exports
• Delete dead code files
• Optimize image formats
• Add missing lazy loading
• Remove unused dependencies

Require Approval (Risky)

• Algorithm refactors
• Dependency swaps
• Code splitting changes
• Build config modifications
• Database index additions

Output: SPEEDRUN.md

# Speedrun Optimization Report

**Date:** YYYY-MM-DD
**Stack:** [detected tech stack]
**Baseline:** Bundle: X MB | Build: Xs | Complexity avg: Y

## Summary
| Hunter | P0 | P1 | P2 | P3 | Total |
|--------|----|----|----|----|-------|
| bundle-hunter | N | N | N | N | N |
| query-hunter | N | N | N | N | N |
| [etc.] | | | | | |

## P0 — Act Now
[Quick wins: high impact, trivial effort]

### [Issue Title]
- **Priority:** P0
- **Hunter:** bundle-hunter
- **Impact:** ~80KB savings
- **Confidence:** HIGH
- **Fix:** [What to do]

## P1 — High Value
[Straightforward fixes worth doing this sprint]

## P2 — Medium Value
[Moderate effort, good returns]

## P3 — Consider Later
[Minor gains, low urgency]

## Verification
| Metric | Before | After | Delta |
|--------|--------|-------|-------|
| Bundle | X MB | Y MB | -Z% |
| Build | Xs | Ys | -Z% |
| Queries/request | N | M | -Z% |

After Optimization

Re-measure all baseline metrics and report the delta. Refuse to claim an improvement without measurement.